Sma scheduling o sa ing ene gy in g id compu ing
A. Fe nández-Mon es
a,
⇑
, L. Gonzalez-Ab il
b
, Juan A. O ega
a
, Lau en Le è e
c
a
Depa amen o de Lenguajes y Sis emas In o má icos, Uni e sidad de Se illa, Spain
b
Depa amen o de Economía Aplicada I, Uni e sidad de Se illa, Spain
c
INRIA RESO – Uni e si é de Lyon – École No male Supé ieu e, F ance
a icle in o
Keywo ds:
Ene gy policies
E ficiency
A anging policies
Simula ion so wa e
abs ac
Ene gy sa ing in ol es wo di ec benefi s: sus ainabili y and cos educ ion. Wi hin he field o In o ma-
ion Technology, clus e s, g ids and da a cen es ep esen he hung ies consume s o ene gy and he e-
o e ene gy (sa ing) policies o hese in as uc u es should be applied in o de o maximize hei
esou ces. I is p o ed in his pape ha app oxima ely 40% o ene gy can be sa ed in a da a cen e i
an adequa e policy is applied. Fu he mo e, a so wa e ool is p esen ed whe e simula ions can be un
and esul s o eal scena ios can be ob ained.
Ó2012 Else ie L d. All igh s ese ed.
1. In oduc ion
In oday’s inc easingly pessimis ic imes no only a e we aced
wi h majo economic issues wo ldwide, bu also wi h hose o sus-
ainabili y. As a pa o sus ainabili y, esea che s a e encou aged
o sa e ene gy in all domains, om In o ma ion Technology (IT)
o anspo . Sa ing ene gy is di ec ly ela ed wi h cu ing cos s
and en i onmen al sus ainabili y. Ene gy e ficiency is he e o e
sough in a wide ange o sys ems om small de ices o la ge-scale
compu ing.
In o ma ion Technology ene gy consump ion ep esen s a me e
3–5% o CO
2
emissions wo ldwide which is simila o ha o a ia-
ion anspo . While appa en ly i ial in quan i y, his usage is
symbolic since IT can g ea ly influence o he indus ial and e-
sea ch domains (Ru h, 2009). As compu ing equi emen s a e e e
g ea e , mic op ocesso manu ac u e s a e doubling he elec ical
e ficiency o compu a ion e e y yea and a hal (Sanchez, Wong,
Be a d, & Koomey, 2011). Ne e heless, ene gy consump ion is s ill
ising despi e hese good esul s, wi h ene gy consump ion o da a
cen es inc easing an a e age o 16.7% o e he las decade (Koo-
mey, 2008).
Some companies, such as Google (Ren, Tune, Moseley, Shi, &
Hund , 2010), a e commi ed o inc easing ene gy e ficiency in
da a cen es and in cloud compu ing. The esea ch communi y
has also been sea ching o imp o emen s in ene gy e ficiency,
whe eas he majo i y o companies ha e ocused hei e o s on
imp o ing acili ies.
The huge amoun o ene gy consumed by g id compu ing p o-
ides jus ifica ion o a s udy in o ene gy-sa ing me hodologies,
ei he om an economic o ecological poin o iew. To his end,
g id ope a ional policies mus be analyzed in o de o be
op imized.
An expe imen al g id o ganiza ion, loca ed in F ance and called
G id’5000, is analyzed in his pape . G id’5000 is a scien ific ins u-
men designed o suppo expe imen -d i en esea ch in all a eas
o compu e science ela ed o pa allel, and la ge-scale and dis ib-
u ed compu ing and ne wo king. I s pu pose is o supply a highly
econfigu able, con ollable and moni o able expe imen al pla -
o m o i s use s.
The es o his pape is s uc u ed as ollows: Sec ion 2includes
a b ie in oduc ion o G id’5000 o ganiza ion and i s cu en en-
e gy consump ion is analyzed. Va ious on–o policies, designed
o sa e ene gy a e p esen ed, and a compa ison be ween cu en
ene gy consump ion and he esul s o each on–o policy a e gi en
in Sec ion 3. The way in which jobs can be scheduled be ween e-
sou ces is shown in Sec ion 4. So wa e de eloped o he es ing
and simula ion is explained in Sec ion 5. Finally, in Sec ion 6, e-
sul s and conclusions a e d awn.
2. G id’5000 a a glance
G id’5000 has been buil upon a ne wo k o dedica ed clus e s.
I is no an ad hoc g id. The in as uc u e o G id’5000 is geog aph-
ically dis ibu ed o e a ious si es, o which he ini ial 9 a e lo-
ca ed in F ance: Bou deaux, G enoble, Lille, Lyon, Nancy, O say,
Rennes, Sophia-An ipolis and Toulouse. Po o Aleg e, in B azil,
and Luxembu g, a e cu en ly being o ficialy included as he 10 h
and 11 h si es, espec i ely. The p ojec began in 2004 as an ini ia-
i e by he F ench Minis y o Educa ion and Resea ch, INRIA,
CNRS, he Uni e si ies o on all he a o emen ioned and se e al e-
gional councils.
The ini ial aim was o achie e 5000 p ocesso s on he pla o m.
This objec i e, e amed a 5000 co es, was eached du ing he
0957-4174/$ - see on ma e Ó2012 Else ie L d. All igh s ese ed.
doi:10.1016/j.eswa.2012.02.115
⇑
Co esponding au ho . Tel.: +34 626215333; ax: +34 954557139.
E-mail add ess: [email p o ec ed] (A. Fe nández-Mon es).
Expe Sys ems wi h Applica ions 39 (2012) 9443–9450
Con en s lis s a ailable a SciVe se ScienceDi ec
Expe Sys ems wi h Applica ions
jou nal homepage: www.else ie .com/loca e/eswa
win e o 2008–2009. On Ma ch 16, 2010, 1569 nodes (5808 co es)
we e in p oduc ion in G id’5000. These si es can now connec each
o he wi hin he same VLAN a 10Gbps hanks o he da k fib e
in as uc u e which connec s hem, in an incomple e g aph
scheme.
G id’5000 allows expe imen s a g id o a clus e le el, which
gua an ees a mo e homogeneous ha dwa e and bandwid h,
al hough g id le el expe imen s a e p e e ed in planning. Each
si e o G id’5000 hos s se e al clus e s, since ha dwa e has been
acqui ed in inc emen al s eps on each si e, whe eby clus e s ha e
been o med a each pu chase. Each clus e is composed o wo
kind o nodes:
Compu e nodes, which cons i u e he base elemen s o a clus e ,
on which compu a ions a e un.
Se ice nodes, which a e dedica ed o hos ing he g id in a-
s uc u e se ices, such as con ol and deploy.
Each node can supply se e al co es, which a e he fines g ains
o esou ce in G id’5000. This means ha i a machine has a mic o-
p ocesso wi h nco es, i o e s n esou ces o he g id.
2.1. Jobs
The pla o m can be used in wo di e en modes: submission
and ese a ion.
Submission: This is used when a job is submi ed o he g id and
he use expec s i o be launched immedia ely. The schedule
decides whe he he job can be un, by aking in o accoun
he occupa ion o cu en esou ces and he agenda o u u e
jobs. Use s usually check hese equi emen s be o e he submis-
sion o a new job h ough a web in e ace ha p esen s he s a e
and he agenda o he g id and i s esou ces.
Rese a ion: This is used when a job is o be launched on he
g id in he u u e. The schedule checks he equi emen s o
ime and esou ces and decides i he ese a ion can be made
o no . Again, use s usually check hese equi emen s be o e
making a ese a ion.
The so wa e used o ask schedule is OAR. This is a esou ce
manage (o ba ch schedule ) o la ge clus e s which allows clus-
e use s o submi o ese e nodes ei he in an in e ac i e o in a
ba ch mode. Job in o ma ion includes submission ime, s a and
s op ime, job iden ifica ion gi en by he manage so wa e, he
owne o he job, and he se o esou ces assigned, which a e going
o un he job. The job in o ma ion includes o he in o ma ion
which is i ele an o his esea ch.
2.2. Resou ces
The G id’5000 pla o m ea u es di e en kinds o machines
depending on he loca ion and he clus e hey belong o, and
when hese machines we e included on he pla o m. Two amilies
can be ound: In el Xeon and AMD Op e on. Each machine o e s i s
CPU co es (usually 2, 4 o 8) o he g id o execu e jobs. Each CPU
co e is called esou ce and each job is ela ed wi h a se o hese
esou ces. Al hough he pe o mance o each esou ce is no iden-
ical, he assump ion ha pe o mances a e e y simila is made,
and hence he e is no e ec i e di e ence be ween unning a job
on one esou ce o ano he . The same assump ion is also made
abou he consump ion o hese esou ces, and he e o e each e-
sou ce uses he same amoun o ene gy.
These assump ions enable job esou ces o be ea anged o
ene gy-sa ing pu poses, wi hou ha ing o conside which ype
o esou ces a job o iginally belonged; hence he e is no di e ence
be ween unning a job on one a se o esou ces o on ano he .
The a ious s a es o he esou ces and hei es ima ed powe
equi ed a e lis ed below:
On. A esou ce is On when i is occupied by a job; he esou ce is
unning he job. A job is usually deployed o e a se o
esou ces. The powe needed is app oxima ely 108 W.
O . A esou ce is O when i has been swi ched o . This means
he esou ce is no occupied wi h any job. The powe needed is
app oxima ely 5 W.
Idle. A esou ce is Idle when i has been swi ched on and wai ing
o new jobs, bu i is no occupied wi h any job. The powe
needed is app oxima ely 50 W.
Boo ing. A esou ce is Boo ing when i is being swi ched on om
O o On. The powe needed is app oxima ely 110 W.
Shu ing. A esou ce is Shu ing when i is being swi ched o
om On o Idle o O . The powe needed is app oxima ely
110 W.
Fig. 1 shows he li e cycle o he esou ces whe e colou s a e
ep esen a i e o u u e figu es, ha is, g een o On, blue o Idle,
ed o Shu ing, g ey o O and yellow o Boo ing. The ime
needed o a s a us change is shown along he edges. No ice ha
s a us changes be ween Idle and On a e immedia e. The imes T
boo -
ing
and T
shu ing
ha e been es ablished o simula ion pu poses as
100 sgs and 10 sgs, espec i ely.
2.3. Pe o mance
Cu en G id’5000 beha iou lea es esou ces Idle while wai ing
o new jobs o un. This policy is he so-called Always On policy
which is he bes o he as sa is ac ion o use s needs, bu he
wo s in e ms o ene gy consump ion. This pape is ocused in
eplacing his cu en policy wi h new policies which a e p esen ed
in he ollowing sec ion.
3. Scheduling ene gy policies
Ene gy policies es ablish he managing o g id esou ces. While
o he esea ch a emp s o educe he makespan (Tseng, Chin, &
Wang, 2009), he policies shown in his wo k aim o desc ibe
and compu e wha o do wi h a esou ce when i finishes he exe-
Fig. 1. Li e cycle o he esou ce.
9444 A. Fe nández-Mon es e al./Expe Sys ems wi h Applica ions 39 (2012) 9443–9450
cu ion o a job. Thus, each ene gy policy decides o ei he lea e e-
sou ces swi ched on o swi ch hem o depending on he pu pose
o he policy. Each ene gy policy is illus a ed wi h a sc eensho o
he G aphical Use In e ace (GUI) whe e he ho izon al axis is he
imeline and he e ical axis ep esen s he esou ces. The ollow-
ing ene gy policies a e implemen ed in he G id’500 Toolbox (Sec-
ion 5):
3.1. Always On
This is he simples ene gy policy, whe eby esou ces a e ne e
swi ched o , unde any condi ion, and he e o e esou ces emain
idle, wai ing o a new job o be un. Cu en ly G id’5000 wo ks un-
de his mode, and hence hese consump ion esul s a e used o
compa ison wi h he esul s o o he ene gy policies in o de o
de e mine how much ene gy can be sa ed. The numbe o imes
esou ces a e swi ched o o on a e always ze o, and he e o e
hei s ess is minimal. Fig. 2 shows he ypical appea ance o e-
sou ces while he G id’5000 Toolbox is unning his ene gy policy.
In his figu e, ou esou ces a e shown, each esou ce is deno ed
by a ow, and ou jobs ha e been ca ied ou .
3.2. Always O
Unde his policy, esou ces a e always swi ches o unde any
condi ion, and he e o e esou ces s a shu ing down a e any
job finishes, and la e hey emain swi ched o . I a new job a -
i es, he assigned esou ces ha e o be boo ed o un ha job. This
boo ing is done wi hin ese a ion limi s, and hus he use is un-
able o make e ec i e use o he esou ces un il hey a e boo ed.
This policy is usually he bes ega ding ene gy consump ion e-
sul s, bu he numbe o imes ha boo ing and shu ing is always
a a maximum, and he e o e he s ess p oduced on he ha dwa e
componen s is he highes , which is no desi able. The ypical
appea ance o esou ces while he G id’5000 Toolbox is unning
unde his ene gy policy is shown in Fig. 3.
3.3. Swi ch O Randomly
This policy swi ches o and andomly lea es he esou ces idle
by ollowing a Be noulli dis ibu ion whose pa ame e is equal o
0.5 when a job finishes. Hence, he imes ha esou ces a e
swi ched o o le idle ends owa ds 50%. Resul s end o be
hal -way be ween hose o he Always O and Always On policies
ega ding he wo kinds o esul s: he imes esou ces a e
swi ched o and ene gy consump ion. The ypical appea ance o
esou ces while he G id’5000 Toolbox is unning unde his en-
e gy policy is shown in Fig. 4.
3.4. Load
Load can be defined as he pe cen age o esou ces ha a e On
among he clus e s o a loca ion. This policy uses his in o ma ion
and ei he swi ches esou ces o i he load, when finishing a job,
is g ea e han he h eshold o lea es he esou ces idle i he load
is less han he h eshold. The h eshold is a pa ame e , anging
om 0 o 1, selec ed om he GUI. The ypical appea ance o e-
sou ces while he G id’5000 Toolbox is unning unde his ene gy
policy wi h 0.6 as i s h eshold pa ame e is shown in Fig. 5.
3.5. Swi ch o T
S
T
S
is defined as he minimum ime ha ensu es ene gy sa ing i a
esou ce is swi ched o be ween wo jobs (O ge ie, Le è e, & Gelas,
2008). T
S
can be compu ed as ollows:
T
S
¼E
s
P
O
d
o
þE
On!O
þE
O !On
P
Idle
P
O
whe e P
O
and P
Idle
e e o he powe consump ion in wa s o a gi-
en esou ce when i is O and Idle, espec i ely. E
On?O
and E
O ?On
e e o he ene gy equi ed in joules o a gi en esou ce o boo o
swi ch i o espec i ely. E
S
is he ene gy sa ed o T
S
seconds. Fi-
nally, d
o
=d
On ?O
+d
O ?On
which is he o al ime a gi en e-
sou ce needs o i o swi ch i sel o and swi ch i sel on.
Fig. 2. Example o Always On policy. Idle (blue), On (g een). (Fo in e p e a ion o
e e ences o colo s in his figu e legend, he eade is e e ed o he web e sion o
his a icle.)
Fig. 3. Example o Always o policy. On (g een), Shu ing Down ( ed), O (g ey),
Boo ing (yellow). (Fo in e p e a ion o e e ences o colo s in his figu e legend, he
eade is e e ed o he web e sion o his a icle.)
Fig. 4. Example o Swi ch o andomly policy. Idle (blue), On (g een), Shu ing down
( ed), O (g ey), Boo ing (yellow). (Fo in e p e a ion o e e ences o colo s in his
figu e legend, he eade is e e ed o he web e sion o his a icle.)
Fig. 5. Example o Load policy. Idle (blue), On (g een), Shu ing Down ( ed), O
(g ey). (Fo in e p e a ion o e e ences o colo s in his figu e legend, he eade is
e e ed o he web e sion o his a icle.)
A. Fe nández-Mon es e al./Expe Sys ems wi h Applica ions 39 (2012) 9443–9450 9445
This ene gy policy uses he agenda o check whe he he subse-
quen submi ed jobs a e going o be un in he g id in less han T
S
.
This policy compu es he numbe o esou ces ha a e needed in a
ime pe iod less han T
S
, and lea es he esou ces o he ecen ly
e mina ed job idle o shu s hem down depending on his compu-
a ion. In his way, he simula o a emp s o minimize boo ing and
shu ing-down cycles when no ene gy can be sa ed. The ypical
appea ance o esou ces while he G id’5000 Toolbox is unning
unde his ene gy policy, whe e T
S
= 130 s, is shown in Fig. 6.
3.6. Exponen ial
The Exponen ial dis ibu ion, deno ed by Exp(k), desc ibes he
ime be ween e en s in a Poisson p ocess, i.e. a p ocess in which
e en s occu con inuously and independen ly a a cons an a e age
a e (1/k). Unde he hypo hesis ha he a i al o new jobs ollows
an Exponen ial dis ibu ion, his ene gy policy a emp s o p edic
he a i al o new jobs. Thus, o compu e he kpa ame e , e e y
ime a job finishes, he mean ime be ween he las jobs is com-
pu ed, and deno ed by
l
. Hence, k=1/
l
by using he me hod o
maximum likelihood. The p obabili y o he a i al o a new job
can hen be compu ed by means o he exponen ial cumula i e
densi y unc ion (cd ) as cd ðT
s
Þ¼1e
T
s
=
l
. The e o e, gi en a
h eshold alue, he ollowing condi ions a e imposed:
i cd ðT
s
Þ>¼ h eshold hen lea e esou ces Idle
i cd ðT
s
Þ< h eshold hen swi ch esou ces O

3.7. Gamma
The Gamma dis ibu ion, deno ed by
C
(h,
j
), is equen ly used
as a p obabili y model o wai ing imes and p esen s a mo e gen-
e al model han he Exponen ial. Unde he hypo hesis ha he a -
i al o new jobs ollows a Gamma dis ibu ion, his ene gy policy
a emp s o p edic he a i al o new jobs. The pa ame e s com-
pu ed e e y ime a job finishes a e:
numbe o esou ces a ailable as esou cesA ailable. These a e
he esou ces ha a e Idle and eady o accep new jobs.
mean esou ces used by las jobs as meanResou ces. To al num-
be o esou ces used by he las jobs is compu ed and di ided
by he numbe o jobs. The numbe o jobs is a selec able win-
dow size.
mean du a ion be ween he p e ious numbe o las jobs as
meanDu a ion. The sum o he du a ion o hese las jobs is com-
pu ed and di ided by he p e ious numbe o las jobs.
 he floo o esou cesA ailable/meanResou ces as z.
The pa ame e s o he Gamma dis ibu ion a e hen es ima ed
as: h=1/meanDu a ion and
j
=z+ 1. Finally he p obabili y o he
a i al o a new job is compu ed by means o he cumula i e den-
si y unc ion (cd ) wi h
cd ðT
s
Þ¼
c
ð
j
;T
s
=hÞ
C
ð
j
Þ
Hence, gi en a h eshold alue, he ollowing condi ions a e
imposed:
i cd ðT
s
Þ>¼ h eshold hen lea e esou ces Idle
i cd ðT
s
Þ< h eshold hen swi ch esou ces O

4. A anging policies
A anging policies es ablish he a anging o he jobs o hei
execu ion. A job can be mo ed om one se o esou ces o an-
o he , o a planned job execu ion can e en be mo ed in ime in o -
de o ake ad an age o esou ces ha a e al eady swi ched on.
Do No hing (DN): does no mo e jobs in ime no om one
esou ce o ano he ; hey a e execu ed as defined in he agenda.
This oge he wi h he ene gy policy Always On o e s he cu -
en G id’5000 beha iou .
Simple Agg ega ion o Jobs (SA): This policy ies o find esou ces
a ailable (Idle) o new jobs. In his way, i a job is assigned o a
se o esou ces which a e O and some o he esou ces a e
a ailable, he ime and he ene gy needed o be swi ched on
can be sa ed. No ice ha his policy does no change s a o
s op imes, and hence i is anspa en o use s.
An example o hese a anging policies can be seen in Fig. 7.
5. G id’5000 Toolbox Simula o
G id’5000 Toolbox
1
eplays he p og ess o he eal g id ega d-
ing he ope a ion o jobs and esou ces. G id’5000 Toolbox is able
o compu e ene gy consump ion o G id’5000, and enables he use
o se up se e al pa ame e s including: (a) simula ion s a - ime,
(b) simula ion s op- ime, (c) loca ion, (d) ene gy policy and (e)
a anging policy. These pa ame e s can be se up h ough he Config-
u a ion ab as shown in Fig. 8.
G id’5000 Toolbox (G id Toolbox, 2011) is a Ja a Desk op appli-
ca ion using lib a ies o: 1. deal wi h ene gy and ime magni udes
(JScience Ma in-Michiello , 2008); 2. communica e wi h RDBMS
Fig. 6. Example o Swi ch o T
S
policy wi h T
S
= 130 seconds. Idle (Blue), On (G een),
Shu ing Down (Red), O (G ey). (Fo in e p e a ion o e e ences o colo s in his
figu e legend, he eade is e e ed o he web e sion o his a icle.)
Fig. 7. Example o a anging policies. i) Be o e jobs ha e been a anged. Job #3 is
abou o s a , assigned o O esou ces. (ii) A e jobs ha e been a anged. Job #3
has been mo ed o a ailable esou ces. No need o boo esou ces.
1
This so wa e can be downloaded and execu ed om he web o he Idin o
esea ch g oup (Idin o , 2011).
9446 A. Fe nández-Mon es e al./Expe Sys ems wi h Applica ions 39 (2012) 9443–9450
(JDBC connec o API Reese, 2000); 3. anno a e da abase en i ies
(Ja a Pe sis ence API JPA Kei h and Schinca iol, 2006; 4. model s a-
is ical dis ibu ions (JSC Ja a S a is ical Classes Be ie, 2002); 5.
logging simula ion in o ma ion (Apache Logging Se ices Log4Ja a
Gup a e al., 2011); 6. w i e esul s in Excel files (JExcel API Khan,
2010); G id’5000 Toolbox includes a module o pa se aw log files
om G id’5000 sys ems and s o es hese files in an s anda d
RDBMS h ough JPA anno a ions. Each log file is ela ed o one
loca ion. The da a ound in hese log files includes: pas jobs, e-
sou ces, machines, clus e s, dead-s a e esou ces, use s, and ela-
ions be ween jobs and esou ces.
The simula o ope a ion is based on an agenda whe e jobs a e
egis e ed and on a lis o esou ces ep esen ing he eal esou ces
a he si es. The simula o s a s que ying he agenda om s a -
ime o s op- ime. Each que y is ela ed o cu en simula ion ime,
and hence he agenda seeks jobs and e en s ha occu a gi en
cu en ime. Once he agenda e u ns new e en s, he simula o
p ocesses hem and changes esou ces s a es as would be needed
o execu ion in he eal wo ld, whils aking in o accoun he pol-
icies selec ed in o de o manage esou ces and jobs. The consumed
ene gy is compu ed s ep by s ep by means o he in o ma ion on
ene gy consump ion o each esou ce and he esou ce s a es de-
ailed in he esou ce lis . The esul s o simula ion execu ions
a e s o ed on a sp eadshee whe e esea che s can find de ails
abou consump ion, he numbe o shu ing and boo ing o e-
sou ces, he compa ison be ween minimal ene gy consumed and
cu en ene gy consumed, e c. Resul s a e also shown in he S a is-
ics ab in a mo e isual way (see Fig. 9).
6. Expe imen a ion
In o de o p esen he esul s, he ecommenda ions o mea-
su ing and epo ing O e all Da a Cen e E ficiency (G een G id,
2011) we e aken in o accoun . Resul s o e e y combina ion o
ene gy and a anging policy summa ize he beha iou o hese pol-
icies o each loca ion and o each ime pe iod selec ed. The com-
pu ed in o ma ion includes:
 o al numbe o boo ings and shu ings,
 o al ene gy consumed
ene gy sa ed as compa ed wi h he ene gy consumed by cu -
en G id’5000 policies (Always Lea e On and Do no A ange
policies)
compa ison be ween he minimal
2
ene gy consumable o an
execu ion and ac ual ene gy consumed by each combina ion o
policies,
compa ison be ween he ene gy consumed by cu en
G id’5000 policies o an execu ion and ac ual ene gy consumed
by each combina ion o policies,
sa ing in ene gy a ained pe shu ing down, which shows he
alidi y o he shu ing down decisions.
6.1. Gene al esul s
Gene al esul s compa e all possible combina ions o ene gy
policies and a anging policies. G id’5000 Toolbox enables
esea che s o un a ba ch o simula ions while defining he pa am-
e e s o each policy. This pape p esen s a summa y o 324 di e -
en simula ions as ollows:
 wo di e en pe iods o six mon hs. F om 1s Janua y, o 30 h
June and om 1s July o 31s Decembe 2008.
 wo a anging policies, Do No hing and Simple Agg ega ion o Jobs
 he se en ene gy policies lis ed in Sec ion 3.
each pa ame e izable ene gy policy has been un wi h a ious
alues o o se e al pa ame e s as ollows:
1. Load policy. Load h eshold pa ame e om 0.0 o 1 by
0.3. A o al o ou scena ios.
2. Exponen ial and Gamma. Th eshold p obabili y pa ame e
om 0.0 o 1 by 0.3, and window size om 2
0
o 2
8
. Hence
he e a e hi y-six di e en scena ios o each policy.
F om he 324 se ups un, he bes ene gy sa e s ha e been se-
lec ed o each policy. Tables 1 and 2 show selec ed esul s o he
wo pe iods.
Wi h espec o he fi s pe iod, he minimal ene gy consumable
is 149,202 kW h o a o al o 74,035 deployed jobs, and he cu en
ene gy consumed by G id’5000 is 217,803 kW h. I can be seen ha
he simples ene gy policy, Always O , is he bes in e ms o en-
e gy sa ing. Howe e , i is he policy which o ces he highes
numbe o powe cycles, and hence he s ess on he ha dwa e is
he g ea es . Load policy wi h 0.9 h eshold e u ns e y simila e-
sul s o ene gy sa ing and numbe o powe cycles.
Fig. 8. Configu a ion ab p esen ing se up pa ame e s o a ba ch o simula ions.
2
The heo e ical minimum ene gy consumed is he sum o he ene gy needed o
un all he jobs o a pe iod. The consump ion by he Idle,Boo ing and Shu ing Down
s a es is no compu ed.
A. Fe nández-Mon es e al./Expe Sys ems wi h Applica ions 39 (2012) 9443–9450 9447

Swi ch O Ts policy is he only policy ha accesses he agenda
o u u e ese a ions be o e deciding wha o do. This ac
explains why he pe cen age o p edic ion success is he g ea es .
I also e u ns e y good esul s in e ms o ene gy sa ing, bu wi h
Fig. 9. S a is ics ab p esen ing esul s o a ba ch o simula ions.
Table 1
Selec ed esul s o Bo deaux om 1s Janua y o 30 h June, so ed acco ding o ene gy consumed.
Ene gy policy A anging policy Boo . + Shu . Ene gy consumed (kW h) Simula ed s Sa ed pe shu .
Name Pa ams Min Cu
Alwz O n/a SA 1,668,900 149,396 100.13 68.59 0.04
Load [0.9] SA 1,569,934 150,304 100.74 69.01 0.04
S.O Ts n/a SA 653,856 150,527 100.89 69.11 0.10
Exp. [0.9, 16] SA 1,047,198 151,733 101.70 69.67 0.06
S.O Rdm n/a SA 890,195 157,645 105.66 72.38 0.07
Gamma [0.9, 1] SA 183,850 186,929 125.29 85.82 0.17
Alwz On n/a SA 0 217,921 146.06 100.05 0.00
Table 2
Selec ed esul s o Bo deaux om 1s July o 31s Decembe , so ed acco ding o ene gy consumed.
Ene gy Policy A anging policy Boo .+ Shu . Ene gy consumed (kW h) Simula ed s Sa ed pe shu .
Name Pa ams Min Cu
Alwz O n/a SA 2,365,598 169,364 100.80 73.61 0.03
S.O Ts n/a SA 1,584,070 169,894 101.11 73.84 0.04
Load [0.9] SA 2,104,456 170,109 101.24 73.93 0.03
S.O Rdm n/a SA 1,321,702 173,323 103.15 75.33 0.04
Exp. [0.9, 1] SA 521,878 175,945 104.71 76.47 0.10
Gamma [0.9, 1] SA 957,784 193,180 114.97 83.96 0.04
Alwz On n/a SA 0 230,093 136.94 100.00 0.00
9448 A. Fe nández-Mon es e al./Expe Sys ems wi h Applica ions 39 (2012) 9443–9450
a much lowe numbe o powe cycles: abou 40% o he numbe o
powe cycles o he Always O policy. This implies a g ea ad an-
age o e Always O and Load-0.9 policies.
S a is ical-based policies, Exponen ial and Gamma, pe o m ea-
sonably well in e ms o he pe cen age o p edic ion success (as
expec ed), pa icula ly he Gamma policy. Ene gy sa ing esul s
o he Exponen ial policy a e a bi wo se han p e ious policies,
and qui e high in e ms o powe cycles. On he o he hand, he
Gamma policy pe o ms modes ly in e ms o ene gy sa ing,
al hough he numbe o powe cycles is he lowes , jus 11% o Al-
ways O , and hence ene gy sa ed pe shu ing down is he g ea es ,
and he s ess imposed on he ha dwa e is he lowes .
Wi h espec o he second pe iod, simila esul s a e ound.
Thus, he minimal ene gy consumable is 168,024 kW h o a o al
o 271,149 deployed jobs, and he cu en ene gy consumed by
G id’5000 is 230,087 kW h. I is wo h no ing ha in his case,
he Exponen ial policy achie es be e esul s han he Gamma pol-
icy. This ac can be explained by aking in o accoun ha he num-
be o jobs deployed du ing he fi s pe iod is one qua e o he
numbe o jobs deployed du ing he fi s pe iod.
6.2. Scheduling a anging policies compa ison
A se o ene gy policies has been selec ed in o de o compa e
esul s be ween he wo a anging policies: Do No hing and Simple
Agg ega ion o Jobs. Resul s a e shown in Table 3.
No ice ha he la e policy, Simple Agg ega ion o Jobs is consis-
en ly he bes since i p o okes ewe powe cycles, sa es mo e
ene gy, and he ene gy sa ed pe shu ing is inc eased in gene al.
The e o e, i jobs a e a anged, e en wi h simple policies, he e-
sul s a e much be e han allowing use s o choose esou ces.
6.3. En i onmen al and Economic esul s
In o de o summa ize ene gy-sa ing esul s, he cos s and CO
2
sa ings a e compu ed o Bo deaux in Table 4. Fo he compu a ion
o hese alues, a p ice o 0.14 eu os pe kW h and a CO
2
gene a ion
o 0.234 kg pe kW h a e conside ed.
No e ha 18,000 eu os and 35 ons o CO
2
pe yea can be sa ed
by implemen ing he bes ene gy sa ing policy. Hence, ex apola -
ing esul s o all 9 loca ions o G id’5000, (Bou deaux, G enoble,
Lille, Lyon, Nancy, O say, Rennes, Sophia-An ipolis and Toulouse),
162,000 eu os and 318 ons o CO
2
pe yea could be sa ed.
In e ms o ene gy, up o 129,254 kW h could be sa ed o Bo -
deaux and by ex apola ing his esul o all 9 loca ions,
1,163,286 kW h. To illus a e how la ge his quan i y o ene gy
eally is, i is equi alen o 78 jou neys o AVE (high speed ail)
ain om Mad id o Ba celona, and i is equi alen o he daily en-
e gy consump ion o 61,314 ci izens in he Eu o zone.
7. Conclusions and Fu u e wo k
Va ious me hodologies o ackling ene gy sa ing in g id com-
pu ing en i onmen s, which could easily be applied o da a cen es
and massi e compu ing en i onmen s a e p esen ed.
We ha e empi ically p o en ha a sui able policy in g id com-
pu ing could sa e a conside able moun o ene gy and educe he
pollu ion o CO
2
in he a mosphe e.
The au ho s a e planning o apply hese echniques o hese
en i onmen s in he u u e, in addi ion o con ibu ing owa ds
he imp o emen o ene gy and a anging policies and hei adjus -
men o new compu ing en i onmen s.
A ecen upg ade o G id’5000 oolbox enables us o e ie e
da a om he clus e loca ed a Cen o In o má ico Cien ífico de And-
alucía (CICA), Spain, and hence u u e wo k will include ex e nal
si es o compa ison wi h G id’5000 si es.
Acknowledgemen
This esea ch is suppo ed by he Spanish Minis y o Science
and Inno a ion R& D p ojec ARTEMISA (TIN2009–14378-C02–01)
Re e ences
Ru h, S. (2009). G een IT mo e han a h ee pe cen solu ion? IEEE In e ne
Compu ing, 13, 74–78.
Table 3
Compa ison o he wo a anging policies compa ison o Bo deaux and a selec ed se o ene gy policies.
Ene gy policy A anging policy Boo .+ shu . Ene gy consumed (kW h) Simula ed s Sa ed pe shu .
Name Pa ams Min Cu
Load [0.9] SA 1,569,934 150,304 100.74 69.01 0.04
Load [0.9] DN 1,561,122 154,141 103.31 70.77 0.04
Load [0.6] SA 902,932 158,205 106.03 72.64 0.07
Load [0.3] SA 299,304 177,839 119.19 81.65 0.13
Load [0.6] DN 892,656 186,406 124.94 85.58 0.04
Load [0.3] DN 292,025 208,141 139.50 95.56 0.03
Load [0.0] SA 650 216,386 145.03 99.35 2.18
Load [0.0] DN 650 216,531 145.13 99.42 1.96
S.O Rdm n/a SA 890,195 157,645 105.66 72.38 0.07
S.O Rdm n/a DN 879,681 183,073 122.70 84.05 0.04
Table 4
Summa y o en i onmen al and economic esul s o Bo deaux.
Ene gy
policy
Ene gy consumed
(kW h)
Sa ed ene gy
(kW h)
Eu os
sa ed
CO
2
kg
sa ed
Ene gy Consumed
(kW h)
Sa ed ene gy
(kW h)
Eu os
sa ed
CO
2
kg
sa ed
Eu os
sa ed
CO
2
kg
sa ed
Alwz S.O 149,396 68,525 9593 18,776 169,364 60,729 8502 16,640 18,096 35,416
Load 150,304 67,617 9466 18,527 170,109 59,984 8398 16,436 17,864 34,963
S.O Ts 150,527 67,394 9435 18,466 169,894 60,198 8428 16,494 17,863 34,960
Exp. 151,733 66,188 9266 18,135 175,945 54,148 7581 14,836 16,847 32,972
S.O Rnd 157,645 60,276 8439 16,516 173,323 56,770 7948 15,555 16,386 32,070
Gamma 186,929 30,992 4339 8492 193,180 36,912 5168 10,114 9507 18,606
Alwz On 217921 0 0 0 230093 0 0 0 0 0
Fi s pe iod Second pe iod To al
A. Fe nández-Mon es e al./Expe Sys ems wi h Applica ions 39 (2012) 9443–9450 9449
Sanchez, M., Wong, H., Be a d, S., & Koomey, J. (2011). Implica ions o his o ical
ends in he elec ical e ficiency o compu ing. IEEE Annals o he His o y o
Compu ing, 33, 46–54.
Koomey, J. G. (2008). Wo ldwide elec ici y used in da a cen e s. En i onmen al
Resea ch Le e s, 3, 034008.
Ren, G., Tune, E., Moseley, T., Shi, Y., & Hund , R.S.R. (2010) Google-Wide p ofiling: a
con inuous p ofiling in as uc u e o da a cen e s, Mic o IEEE 30, 65–79.
Tseng, L.-Y., Chin, Y.-H., & Wang, S.-C. (2009). A minimized makespan schedule
wi h mul iple ac o s o G id compu ing sys ems. Expe Sys ems wi h
Applica ions, 36, 11118–11130.
O ge ie, A.-C., Le è e, L., Gelas, J.-P.(2008). Sa e Wa s in you g id: g een s a egies
o ene gy-awa e amewo k in la ge scale dis ibu ed sys ems. In: 2008 14 h
IEEE In e na ional Con e ence on Pa allel and Dis ibu ed Sys ems, pp.171–178.
Idin o , h p://madei a.lsi.us.es/G upoIdin o , 2011.
G id Toolbox, h p://madei a.lsi.us.es/p oyec osWeb/g id.php, 2011.
Ma in-Michiello , S. (2008).JScience, a gene al scien ific API in Ja a, P ac ice.
Reese, G. (2000). Da abase P og amming wi h JDBC and Ja a (2nd ed.). O’Reilly &
Associa es, Inc..
Kei h, M., & Schinca iol, M. (2006). P o EJB 3: Ja a Pe sis ence API (P o). Be kely, CA,
USA: Ap ess.
Be ie, A. (2002). Ja a applica ions o eaching s a is ics. MSOR Connec ions, 2,
78–81.
Gup a, S. (2011). P o Apache Log4j (2nd ed.). Ap ess.
Khan, A. (2010). JExcel: A Ja a lib a y o eading/w i ing Excel, Accessed Online a
h p://jexcelapi.sou ce o ge.ne /.
G een G id, T. (2011). Recommenda ions o measu ing and epo ing o e all da a
cen e e ficiency Ve sion 2 – Measu ing PUE o Da a Cen e s.
9450 A. Fe nández-Mon es e al./Expe Sys ems wi h Applica ions 39 (2012) 9443–9450